Deep Dive Architecture

Technical specification for the automated daily intelligence briefing system.

System Overview

┌─────────────┬─────────────┬─────────────┬─────────────┬─────────────┐
│   Phase 1   │   Phase 2   │   Phase 3   │   Phase 4   │   Phase 5   │
│  Aggregate  │   Filter    │  Synthesize │    TTS      │   Deliver   │
├─────────────┼─────────────┼─────────────┼─────────────┼─────────────┤
│ arXiv RSS   │ Chroma DB   │ Claude/GPT  │   Piper     │  Telegram   │
│ Lab Blogs   │ Embeddings  │   Prompt    │   (local)   │   Voice     │
└─────────────┴─────────────┴─────────────┴─────────────┴─────────────┘

Data Flow

Aggregation: Fetch from arXiv + lab blogs
Relevance: Score against Hermes context via embeddings
Synthesis: LLM generates structured briefing
TTS: Piper converts to audio (Opus)
Delivery: Telegram voice message

Source Coverage

Source	Method	Frequency
arXiv cs.AI	RSS	Daily
arXiv cs.CL	RSS	Daily
arXiv cs.LG	RSS	Daily
OpenAI Blog	RSS	Weekly
Anthropic	RSS	Weekly
DeepMind	Scraper	Weekly

Relevance Scoring

Keyword Layer: Match against 20+ Hermes keywords Embedding Layer: all-MiniLM-L6-v2 + Chroma DB Composite: 0.3 * keyword_score + 0.7 * embedding_score

TTS Pipeline

Engine: Piper (en_US-lessac-medium)
Speed: ~1.5x realtime on CPU
Format: WAV → FFmpeg → Opus (24kbps)
Sovereign: Fully local, zero API cost

Cron Integration

job:
  name: deep-dive-daily
  schedule: "0 6 * * *"
  command: python3 orchestrator.py --cron

On-Demand

python3 orchestrator.py        # Full run
python3 orchestrator.py --dry-run    # No delivery
python3 orchestrator.py --skip-tts   # Text only

Acceptance Criteria

Criterion	Status
Zero manual copy-paste	✅ Automated
Daily 6 AM delivery	✅ Cron ready
arXiv + labs coverage	✅ RSS + scraper
Hermes relevance filter	✅ Embeddings
Written briefing	✅ LLM synthesis
Audio via TTS	✅ Piper pipeline
Telegram delivery	✅ Voice API
On-demand command	✅ CLI flags

Epic: #830 | Status: Architecture Complete

2.7 KiB Raw Blame History